The imbalance and concept drift problems in data streams become more complex in multi-class environment, and extreme imbalance and variation in class ratio may also exist. To tackle the above problems, Hybrid Sampling and Dynamic Weighted-based classification method for Multi-class Imbalanced data stream (HSDW-MI) is proposed. The HSDW-MI algorithm deals with imbalance and concept drift problems through the hybrid sampling and dynamic weighting phases, respectively. In the hybrid sampling phase, adaptive spectral clustering is proposed to sample the data after clustering, which can maintain the original data distribution; then the sample safety factor is used to determine the samples to be sampled for each class; the safe samples are oversampled and the unsafe samples are under-sampled in each cluster. If the data stream is extremely imbalanced, the sample storage pool is used to extract samples with a high safety factor to add to the data stream. In the dynamic weighting phase, a dynamic weighting method based on the G-mean value is proposed. The G-mean values are used as the weights of each base classifier in the ensemble and the ensemble is dynamically updated during the processing of the data stream to accommodate the occurrence of concept drift. Experiments were conducted with LB, OAUE, ARF, BOLE, MUOB, MOOD, CALMID, and the proposed HSDW-MI on 10 multi-class synthetic data streams with different class ratios and concept drifts and 3 real multi-class imbalanced streams with unknown drifts, and the results show that the proposed HSDW-MI has better classification capabilities and performs more consistently compared to all other algorithms.

Link prediction is a crucial problem in the analysis of graph-structured data, and graph neural networks (GNNs) have proven to be effective in addressing this problem. However, the computational and temporal costs associated with large-scale graphs remain a concern. This study introduces a novel method for link prediction called Sparse Subgraph Prediction Based on Adaptive Attention (SSP-AA). The method generates sparse subgraphs and utilizes Graph SAmple and aggreGatE (GraphSAGE) for prediction, aiming to reduce computation and time costs while providing a foundation for future exploration of large-scale graphs. Certain key issues in GraphSAGE are addressed by integrating an adaptive attention mechanism and a jumping knowledge module into the model. To address the issue of adaptive weight distribution in GraphSAGE, an aggregation function is employed, which is based on the attention mechanism. This modification enables the model to distribute weights adaptively among neighboring nodes, significantly improving its ability to capture node relationships. Furthermore, to tackle the common issue of over-smoothing in GNNs, a jumping knowledge module is integrated, enabling information sharing across different layers and providing the model with the flexibility to select the appropriate representation depth based on the specific situation. By enhancing the quality of node representations, SSP-AA further boosts the performance of GraphSAGE in various prediction tasks involving graph-structured data.

Considering the complexity of entity pair relations and the information contained in the target neighborhood in few-shot knowledge graphs (KG), existing few-shot KG completion methods generally suffer from insufficient relation representation learning capabilities and neglecting the contextual semantics of entities. To tackle the above problems, we propose a Few-shot Relation Learning-based Knowledge Graph Completion model (FRL-KGC). First, a gating mechanism is introduced during the aggregation of higher-order neighborhoods of entities in formation, enriching the central entity representation while reducing the adverse effects of noisy neighbors. Second, during the relation representation learning stage, a more accurate relation representation is learned by using the correlation between entity pairs in the reference set. Finally, an LSTM structure is incorporated into the Transformer learner to enhance its ability to learn the contextual semantics of entities and relations and predict new factual knowledge. We conducted comparative experiments on the publicly available NELL-One and Wiki-One datasets, comparing FRL-KGC with six few-shot knowledge graph completion models and five traditional knowledge graph completion models for five-shot link prediction. The results showed that FRL-KGC outperformed all comparison models in terms of MRR, Hits@10, Hits@5, and Hits@1 metrics.